Traffic Generated by Mixed-Use
Developments—Six-Region Study Using
Consistent Built Environmental Measures
Reid Ewing1; Michael Greenwald2; Ming Zhang3; Jerry Walters4; Mark Feldman5;
Robert Cervero6; Lawrence Frank7; and John Thomas8
Abstract: Current methods of traffic impact analysis, which rely on rates and adjustments from the Institute of Transportation Engineers,
are believed to understate the traffic benefits of mixed-use developments (MXDs), leading to higher exactions and development fees than
necessary and discouraging otherwise desirable developments. The purpose of this study is to create new methodology for more accurately
predicting the traffic impacts of MXDs. Standard protocols were used to identify and generate data sets for MXDs in six large and diverse
metropolitan regions. Data from household travel surveys and geographic information system (GIS) databases were pooled for these MXDs,
and travel and built environmental variables were consistently defined across regions. Hierarchical modeling was used to estimate models for
internal capture of trips within MXDs, walking and transit use on external trips, and trip length for external automobile trips. MXDs with
diverse activities on-site are shown to capture a large share of trips internally, reducing their traffic impacts relative to conventional suburban
developments. Smaller MXDs in walkable areas with good transit access generate significant shares of walk and transit trips, thus also
mitigating traffic impacts. Centrally located MXDs, small and large, generate shorter vehicle trips, which reduces their impacts relative
to outlying developments. DOI: 10.1061/(ASCE)UP.1943-5444.0000068. © 2011 American Society of Civil Engineers.
CE Database subject headings: Traffic management; Assessment; Environmental issues.
Author keywords: Mixed-use development; Trip generation; Internal capture; Traffic impact assessment.
Introduction
Mixed-use development (MXD) is a signature feature of smart
growth, New Urbanism, and other contemporary land-use
movements aimed at reducing the private automobile’s dominance
1
Professor, Dept. of City and Metropolitan Planning, Univ. of Utah, 375
S. 1530, E. Salt Lake City, UT 84103 (corresponding author). E-mail:
[email protected]
2
Senior Transportation Planner, Lane Council of Governments, 859
Willamette St., Suite 500, Eugene, OR 97401; formerly, Lead Modeler
and Analyst/GIS Specialist, Urban Design 4 Health, Inc., P.O. Box
85508, Seattle, WA 98145. E-mail: [email protected]
3
Associate Professor, Community and Regional Planning, Univ. of
Texas at Austin, 1 University Station, B7500, Austin, TX 78712. E-mail:
[email protected]
4
Principal and Chief Technical Officer, Fehr & Peers, 100 Pringle Ave.,
Suite 600, Walnut Creek, CA 94596. E-mail: [email protected]
5
Transportation Engineer, Fehr & Peers, 100 Pringle Ave., Suite 600,
Walnut Creek, CA 94596. E-mail: [email protected]
6
Professor, Dept. of City and Regional Planning, Univ. of California,
Berkeley, 228 Wurster Hall, Berkeley, CA 94720-1850. E-mail: robertc@
berkeley.edu
7
Associate Professor, School of Environmental Health, Univ. of British
Columbia, 3rd Floor—2206 East Mall. Vancouver, BC V6T 1Z3 Canada.
E-mail: [email protected]
8
Policy Analyst, Office of Sustainable Communities, US EPA, 1200
Pennsylvania Ave. NW Mailcode 1807T, Washington, DC 20816. E-mail:
[email protected]
Note. This manuscript was submitted on December 11, 2009; approved
on October 11, 2010; published online on October 19, 2010. Discussion
period open until February 1, 2012; separate discussions must be submitted
for individual papers. This paper is part of the Journal of Urban Planning
and Development, Vol. 137, No. 3, September 1, 2011. ©ASCE, ISSN
0733-9488/2011/3-248–261/$25.00.
in suburban America. By putting offices, shops, restaurants,
residences, and other codependent activities in close proximity
to each other, MXD shortens trips and thus allows what might
otherwise be external car trips to become internal walk, bike, or
transit trips. This in turn can reduce the vehicle miles generated
by an MXD relative to what it would be if the same activities were
separated in single-use developments. Fewer vehicle miles traveled
(VMT) not only relieves traffic congestion but also reduces greenhouse gas (GHG) emissions, air pollution, and fuel consumption.
MXDs are also promoted for their supply side benefits, such as
possibilities for shared parking and economizing on roadway
and related infrastructure expenditures (because peak travel periods
often differ between offices, retail, and other uses, enabling investments to be descaled) (Cervero 1988).
A diverse group of stakeholders has a vested interest in the
traffic impacts of MXDs. The replacement of off-site car trips with
on-site walking or cycling or (for larger mixed-use sites) on-site
transit or driving matters to developers who want smooth-flowing
traffic conditions to help market their projects, to communities that
want to keep existing residents safe from traffic impacts, and to
traffic engineers whose very profession is devoted to facilitating
traffic flows but often harbor some skepticism about the traffic
benefits of MXDs.
Accurately estimating the proportion of trips captured internally
by MXDs is vitally important if communities are to accurately
assess their traffic impacts and reward such projects through lower
exactions and development fees or expedited project approvals.
However, lacking a reliable methodology for adjusting trip generation estimates, communities face a dilemma when assessing MXD
proposals: do they err on the conservative side by downplaying
internal capture and thereby potentially discourage worthwhile
projects, or err on the liberal side and risk unmitigated traffic
248 / JOURNAL OF URBAN PLANNING AND DEVELOPMENT © ASCE / SEPTEMBER 2011
Downloaded 14 Sep 2011 to 192.245.136.3. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
impacts? Often, the “do no harm” sentiment prevails: when in doubt,
go with conventional practices, which, with MXD proposals, typically means only a small downward adjustment in estimated trips, if
any adjustment at all.
In addition to getting internal capture estimates right, accurate
assessments of MXD projects also depend on estimating the share
of external trips served by alternative modes (e.g., transit and walking). These must also be subtracted from nominal trip generation
rates to estimate the net impacts of MXDs on traffic and VMT.
Community acceptance depends on whether a proposed MXD is
perceived as a good neighbor. Exaggerated estimates of a project’s
traffic generation can heighten concerns about congestion, community image and character, and even public health and safety. A
nimby backlash can add substantially to the time and expense
of securing project approval, and can result in the project being
scaled back to a level at which elected officials feel that the trip
generation is more acceptable. However, the market demand for
the development that is disallowed does not vanish and more often
than not ends up in another location, often at a lower density and in
a less mixed-use configuration. The end result can be more traffic
and higher overall VMT than if the original MXD proposal had
been approved.
Traffic generation estimates have supply-side impacts, affecting
project design and cost. This includes the most obvious components, such as street widths, parking supply, and access point
design, and secondary effects on the design and cost of ancillary
infrastructure like storm water drainage systems. Within constrained sites, overdesign of traffic elements can limit the space
available for revenue-producing land uses and increase other development costs; forcing, for example, the rerouting of utilities to
accommodate traffic-handling infrastructure.
Statutory laws also place pressure on getting the traffic estimates
for MXDs right. The formal National Environmental Policy Act
(NEPA) process and state-level environmental reviews rely on traffic generation estimates to assess impacts and dictate a project’s
mitigation measures. Estimates that exaggerate negative impacts
increase the likelihood a project will be judged as a significant
threat to environmental quality. This leads to more rigorous analysis and reporting of a wide array of potential secondary impacts,
such as the growth inducing effects of the additional required traffic
capacity. It also prompts a more protracted discussion of community concerns through formal public involvement, document review, comment/response periods, and certification protocols. In
addition, it can incorrectly trigger a review of other impacts such
as noise, air quality, energy use, and greenhouse gas emissions.
Development fee programs rely heavily on traffic generation
estimates. As the most comprehensive and widely used reference
on the subject, the Trip Generation report of the Institute of Transportation Engineers (ITE) (2008) has become the principal data
source for setting transportation development fee rates. Most cities,
counties, and regional agencies opt for uniformity rather than
accuracy in this regard. In the interest of standardization of assumptions and approach, many jurisdictions rely on the numbers in
Trip Generation to quantify traffic impacts and mitigation fee
schedules. The unquestioning use of the ITE report can unreasonably jeopardize a MXD project’s approval, financial feasibility, and
design quality.
vehicle and whose origins or destinations lie outside the development. Quoting the report: “Data were primarily collected at suburban localities with little or no transit service, nearby pedestrian
amenities, or travel demand management (TDM) programs.” Recognizing but not resolving this limitation, Trip Generation advises:
“At specific sites, the user may want to modify the trip generation
rates presented in this document to reflect the presence of public
transportation service, ridesharing or other TDM measures, enhanced pedestrian and bicycle trip-making opportunities, or other
special characteristics of the site or surrounding area.” Unfortunately, the desire among traffic engineers for standardization and
substantial documented evidence prevents them from taking this
advice in the vast majority of cases.
Even setting aside the variety and complexity of mixed-use developments, the reliability of Trip Generation for evaluating simple
single-use developments is less than one might assume. For example, for the most widely studied land-use category within the report,
single-family residential, the average ITE Trip Generation daily
rate is 9.6 vehicle trips per dwelling unit, but the standard deviation
is 3.7, almost 40% of the mean. Even for this most uniform of landuse types, when described in terms of a single descriptive variable
(number of dwelling units), the 9.6 ITE trip generation figure used
in impact study guidelines and fee ordinances throughout the
United States is just a midpoint in a standard deviation range of
5.9 to 13.3 vehicle trips per dwelling. The standard deviations
for other common and well-studied land-use types (e.g., office
buildings and shopping centers) are at least 50% of the mean
values. Clearly, trip generation estimation is far from a precise
science.
Professor Donald Shoup of UCLA has been particularly critical
of Trip Generation for the following reasons: the false precision
with which average trip generation rates are presented, the small
samples upon which average rates and regression equations are
based, the insignificance of regression coefficients and constants,
the implicit assumption that trip generation increases with building
size, and the disregard of factors other than building size in the
regression analyses. Pointing to the need to represent land-use interactions more carefully, Shoup remarks, “Floor area is only one
among many factors that influence vehicle trips at a site, and we
should not expect floor area or any other single variable to accurately predict the number of vehicle trips at any site or land-use”
(Shoup 2003).
As an indication of just how far off the mark ITE rates may land,
the Transit Cooperative Research Program (TCRP) recently funded
a study of the trip generating characteristics of transit-oriented developments (TODs) (Cervero and Arrington 2008). The aim was to
seed the ITE Trip Generation report with original and reliable trip
generation data for one important TOD land-use—housing—with
the expectation that other TOD land uses will be added later. For all
TOD housing projects studied, weekday vehicle trip rates were
considerably below the ITE average rate for similar uses. Taking
the weighted average across the 17 case-study projects, TOD housing projects, which included both multifamily and single-family
residential units, generated approximately 44% fewer vehicle trips
than predicted by the ITE report (3.8 trips per dwelling unit for
TOD housing versus 6.7 trips per dwelling unit by ITE estimates
for the site-specific mixes of multi- and single-family residences).
Conventional Traffic Impact Analysis
ITE Method for MXDs
Virtually all traffic impact analyses rely on the ITE Trip Generation
report (2008). The ITE rates are largely representative of individual,
single-use suburban developments whose trips are by private
For mixed-use development projects, Chapter 7 of Trip Generation
Handbook (2004) outlines a procedure for estimating the proportion of trips that remain within the development (i.e., the internal
JOURNAL OF URBAN PLANNING AND DEVELOPMENT © ASCE / SEPTEMBER 2011 / 249
Downloaded 14 Sep 2011 to 192.245.136.3. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
capture rate), and hence place no strain on the external street
network. An ITE member survey found that nearly two-thirds of
practitioners estimate internal capture rates using this procedure.
The procedure works as follows:
1. The analyst determines the amounts of different land-use
types (residential, retail, and office) contained within the
development.
2. These amounts are multiplied by ITE’s per-unit trip generation
rates to obtain a preliminary estimate of the number of vehicle
trips generated by the site. This preliminary estimate is what
the site would be expected to generate if there were no interactions among the on-site uses.
3. The generated trips are then reduced by a certain percentage to
account for internal-capture of trips within MXDs. The reductions are based on lookup tables. The share of internal trips
from the appropriate lookup table is multiplied by total
numbers of trips generated by a given use to obtain an initial
estimate of internal trips for each producing use and attracting
use.
4. For each pair of land uses, productions and attractions are
reconciled such that the number of internal trips produced by
one use just equals the number attracted by the other use. The
lesser of the two estimates of internal trips constrains the number of internal trips generated by the other use.
Strengths of the Current ITE Method
From the viewpoint of the practicing engineer, the ITE internal
capture methodology has some important advantages:
1. It seems objective. Two analysts given the same data will arrive
at exactly the same result. There is no room for negotiation or
interpretation (and therefore no reason to pressure the analyst
into skewing the results in a predetermined direction).
2. It seems logical. Most engineers readily accept the idea that the
degree of internalization will be determined by how well the
productions and attractions match for each trip purpose.
3. It is fast. With a spreadsheet template, an analyst can input the
data and have an answer in a matter of minutes.
Viewed another way, any new methodology that lacks these
qualities may not find wide acceptance within the engineering
community.
Weaknesses of the Current Method
The ITE methodology also has major shortcomings:
1. The two lookup tables are based on data for a “limited number
of multiuse sites in Florida” (specifically, three sites analyzed
by the Florida Department of Transportation, ITE 2004). The
accuracy of forecasts is thus dependent on how closely the site
being analyzed matches the sites used in the tables’ creation.
The fact that the data are drawn from the suburbs of Florida
casts doubt on the applicability to other parts of the country.
The handbook acknowledges this problem and instructs the
analyst to find analogous sites locally and collect his own data
to produce locally valid lookup tables.
2. The land-use types and adjustments embodied in the lookup
tables are limited to the three uses: residential, retail, and
offices. The traffic impacts of other mixed uses cannot be
assessed.
3. The scale of development is disregarded. Clearly, a large site
with many productions and attractions is more likely to produce matches than a small site, and the lookup tables for
large sites should have higher cell percentages than the tables
for small sites. Development scale was the most significant
influence on internal capture rates in a study of South Florida
MXDs, and more than half of all trips were found to be
internalized by community-scale MXDs, far in excess of
any rate obtainable with the handbook method (Ewing et al.
2001).
4. The land-use context of development projects is ignored. Common sense and the literature tell us that projects in remote
locations are more likely to capture trips on-site than those
surrounded by competing trip attractions. For MXDs in South
Florida, the second most important determinant of internal capture rates was accessibility to the rest of the region (second
after the scale of development). Conversely, projects in areas
of high accessibility are more likely to generate walk trips to
external destinations.
5. The possibility of mode shifts for well-integrated, transitserved sites is not explicitly considered. This may not bias
results for free-standing sites, but infill projects within an
urban context may capture few trips internally but still have
significant vehicle trip reductions relative to the ITE rates.
6. The length of external private vehicle trips is not considered.
The ITE methodology deals with trip generation, not traffic
generation. Clearly, in terms of roadway congestion, emissions, and other impacts, a 10-mile trip has a greater impact
than a 5-mile trip. Trip distribution is an ad hoc process under
the ITE methodology.
Parallel Efforts
There would seem to be two ways to refine ITE trip generation
estimates of MXDs. One, a bottom-up approach, would add to
the paltry set of development projects that currently constitute
the ITE database on MXDs, analyze in detail this larger sample’s
trip-making characteristics, and then derive a set of more complete
adjustments to ITE trip rates. In an effort parallel to our own, the
National Cooperative Highway Research Program (NCHRP) is
taking this approach in NCHRP Project 8-51, “Enhancing Internal
Trip Capture Estimation for Mixed-Use Developments.” Adding
four sites to the three that currently form the basis for internal
capture calculations in ITE’s Trip Generation Handbook (2004),
the project has developed an estimation procedure that includes
a proximity adjustment to account for project size and layout.
A second approach, more top down in nature, would assemble
enough data on MXDs to estimate statistical models of traffic
generation in terms of standard built environmental variables,
the so-called “D” variables of density, diversity, design, destination
accessibility, distance to transit, and development scale. Taking this
approach, Ewing et al. (2001) modeled internal capture rates for 20
mixed-use communities in South Florida. For the 20 communities,
internal capture rates ranged from 0 to 57% of all trip ends generated by the community.
To explain this variation, internal capture rates were modeled in
terms of land-use and accessibility measures. The variable that
proved most strongly related to internal capture was neither landuse mix nor density, but the size of the community itself. The two
communities with the highest internal capture rates, Wellington and
Weston, also are the largest, each having more than 30,000 residents and 5,000 jobs. Indeed, these two communities are large
enough to have incorporated as their own small cities. The second
most important variable was regional accessibility, which was
inversely related to internal capture rates. Both of these communities are on the western edge of development in Southeast Florida,
far from other population centers.
Owing to size and inaccessibility, these communities capture a
much higher percentage of trips internally than, for example, the
higher density and better-mixed Miami Lakes. However, Miami
250 / JOURNAL OF URBAN PLANNING AND DEVELOPMENT © ASCE / SEPTEMBER 2011
Downloaded 14 Sep 2011 to 192.245.136.3. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
Lakes doubtless generates shorter auto trips and many more walk,
bicycle, and transit trips than the other two. Its overall impact on the
regional road network is almost certainly less.
The validity and reliability of Ewing et al.’s results are limited by
the small sample, limited geography coverage, and small number of
built environmental variables. The present study improves on the
earlier study by, in this order: (1) pooling travel and built environmental data for 239 MXDs in six diverse regions; (2) consistently
defining travel outcomes and built environmental variables for
these MXDs and regions; (3) estimating models of internal capture,
external walk and transit choice, and external private vehicle trip
length using hierarchical modeling methods; and (4) validating
the results through comparison to traffic counts at an independent
set of mixed-use sites in various parts of the United States.
Conceptual Framework
In travel research, urban development patterns have come to be
characterized by “D” variables. The original “three Ds,” coined
by Cervero and Kockelman (1997), are density, diversity, and
design. Additional Ds have been labeled since then: destination
accessibility, distance to transit, and demographics (Ewing and
Cervero 2001, 2010). An additional D variable is relevant to this
analysis: development scale.
The theory of rational consumer choice underlies this study. It is
well articulated elsewhere (for example, Crane 1996; Boarnet and
Crane 2001; Cervero 2002; Zhang 2004; Cao et al. 2009). Travel
to/from MXDs is conceived as a series of choices that depend on
the D variables (see conceptual framework in Fig. 1). The choices
relate directly to the methodology this paper is proposing to adjust
ITE trip generation rates downward.
The first adjustment to ITE rates is for trips that remain within
the development. Destination choice is conceived as dichotomous.
A traveler may choose a destination within the development, or a
destination outside the development. Internal trips are treated as
100% deductions from ITE trip generation rates.
Then, for trips that leave the development, adjustments are
made for walking and transit use. Mode choices are conceived
as dichotomous. For external trips, a traveler may choose to walk
or not. Likewise, the traveler may choose to use transit or not.
Walking and transit use may be treated as 100% deductions from
ITE trip generation rates, or may be treated as partial offsets. It is
reasonable to assume that transit trips substitute for automobile
trips, but walk trips may supplement or substitute for automobile
trips. The study team plans to propose substitution rates based on a
review of literature.
Finally, for external personal vehicle trips, the traveler chooses
a destination. This destination may be near or far. This outcome
variable is continuous rather than dichotomous.
The D variables in Fig. 1 are characteristics of travelers, MXDs,
and regions, as defined in the following. The D variables determine,
moderate, mediate, and confound travel decisions.
Sample Selection
A main criterion for inclusion of regions in this study was data
availability. Regions had to offer:
1. Regional household travel surveys with XY coordinates for
trip ends, so we could distinguish trips to, from, and within
small MXDs; and
2. Land use databases at the parcel level with detailed land-use
classifications, so we could study land-use intensity and mix
down to the parcel level.
Most U.S. regions fall short on one or both counts. Although
nearly all metropolitan planning organizations (MPOs) have
conducted regional household travel surveys as the basis for the
calibration of regional travel demand models, most have geocoded
trip ends only at the relatively coarse geography of traffic analysis
zones. Likewise, although most MPOs have historical land-use
databases that are used in model calibration, these too provide data
only for the relatively coarse geography of traffic analysis zones.
Traffic analysis zones vary in size from region to region, but as a
general rule, are equivalent to census block groups. They are large
relative to many MXDs, and in any event, will ordinarily not
coincide with MXD boundaries.
Thirteen regional household travel databases were identified
that met the first criterion. This was narrowed down to six regions
based on the availability of parcel-level land-use data and the
experience of planning researchers who had worked with these data
sets.
All six travel databases were derived from large-scale household
travel diary surveys. All allowed the writers to classify trips by
purpose and mode of travel. All allowed us to control for socioeconomic characteristics of travelers that may confound interactions
between the built environment and travel. All had already been
linked to built environmental databases. Although the specific
variables differed somewhat from database to database, it was possible to reconcile differences and specify equivalent models. Also,
although the years of the surveys differed, it was possible to control
for these and other fixed effects at the regional level in a three-level
hierarchical model.
Identifying MXDs
The ITE definition of multiuse development was modified to create
a generic definition of MXD that would encompass many existing
areas with interconnected, mixed land-use patterns:
“A mixed-use development or district consists of two or more
land uses between which trips can be made using local streets,
without having to use major streets. The uses may include
residential, retail, office, and/or entertainment. There may
be walk trips between the uses.”
Fig. 1. Traffic impact adjustments
To identify MXDs in the six study regions at the dates of the
most recent regional household travel surveys, the team used a
JOURNAL OF URBAN PLANNING AND DEVELOPMENT © ASCE / SEPTEMBER 2011 / 251
Downloaded 14 Sep 2011 to 192.245.136.3. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
bottom-up, expert-based process in which planners for the different
localities were queried about MXDs within their boundaries. Using
this approach, a definition of an MXD was read to local planners
over the phone, and they were asked to name, identify the boundaries, and list the uses contained within such areas. In two of the
six regions, local traffic engineers and ITE members were asked to
review the selected sites to confirm that they met the criteria normally applied by practitioners to identify mixed-use developments.
Variables
Final Samples
A total of 239 MXDs were identified, ranging from a low of 24 in
Atlanta to a high of 59 in Boston. Site characteristics ranged from
compact infill sites near the region’s core to low-rise freeway oriented developments. The 239 survey sites varied in population and
employment densities, mix of jobs and housing, presence or absence of transit, and location within the region. The sites ranged
in size from less than five acres to over 2,000 acres, and over
15,000 residents and employees.
Sample statistics are shown in Table 1. The regions that contribute modest numbers of trip ends to the sample still add statistical
power. The importance of Boston, Houston, and Sacramento lies in
the number of MXDs each contributes, not in the number of trip
ends. Also, the inclusion of the three regions doubles the number of
regions in the sample. In a hierarchical analysis, statistical power is
limited by the number of degrees of freedom at each level of analysis. There are ample cases at Level 1, the trip end level, but a shortage of cases at Level 2, the MXD level, and a severe shortage at
Level 3, the regional level.
RiverPlace, a classic MXD just south of downtown Portland, is
one of the 239 MXDs in our sample (Fig. 2). Its internal capture
rate is a surprisingly high 36%. Of the external trips, 14% are made
Table 1. Sample Statistics
Survey
year
Atlanta
Boston
Houston
Portland
Sacramento
Seattle
Total
2001
1991
1995
1994
2000
1999
by walking and 9% by transit. Its external auto trips average
7.7 miles. According to the National Household Travel Survey
of 2009, 14% of Portland’s trips are by walking, and 2% are by
transit. The average vehicle trip length in the Portland Consolidated
Metropolitan Statistical Area is 8.9 miles. On balance, the traffic
impact of RiverPlace is a fraction of the regional average.
MXDs
24
59
34
53
25
44
239
Mean acreage
Total
Mean trip
per MXD
trip ends ends per MXD
287
175
401
116
179
207
211
6,167
3,578
1,584
6,146
2,487
15,915
35,877
Fig. 2. (Color) RiverPlace at eye level
257
61
47
116
99
362
150
In this study, all seven types of D variables were measured and used
to predict the travel characteristics of MXDs (Tables 2 and 3). The
richness of the data sets varies from region to region. Portland and
Atlanta have the most complete data sets. Houston and Sacramento
have the least complete data sets. Sidewalk data, for example, are
only available for Atlanta and Portland. Floor area ratios are only
available for Atlanta, Portland, and Seattle. Measures of job mix are
only available for Atlanta, Boston, Portland, and Seattle.
To maximize the sample of MXDs, the writers decided to limit
our analysis to built environmental variables available for all six
regions. This also simplifies the use of the resulting models by
practitioners, who may have incomplete data on their projects or
their surroundings.
There is great variation in internal capture rates from MXD to
MXD and region to region. Across regions, average internal capture rates vary from a low of 15.9% in Portland to a high of 31.1%
for Houston (Table 4). The high rate for Houston may reflect the
fact that Houston’s MXDs are, in general, larger and more remotely
located than those in other regions. Many are standalone masterplanned communities.
In all household travel surveys, automobile, walk, and transit
(bus or rail, where available) are identified as separate modes of
travel. Bicycle is as well, but samples are too small to be reliably
analyzed. In all regions, the dominant mode for external trips to/
from MXDs is the automobile (“private motor vehicle”). The
essential choices facing travelers are to walk or use an automobile,
or to take transit or use an automobile. For external trips, average
mode shares by walking and transit combined vary from a low of
3.3% for Sacramento to a high of 28.4% for Boston (Table 4).
Of the 35,877 trip ends generated by these MXDs, 6,378
(17.8%) involved trips within the mixed-use site, another 2,099
(5.9%) involved trips entering or leaving the site via walking,
and another 1,995 (5.6%) involved trips entering or leaving via
transit. Thus, on average, a total of 29% of the trip ends generated
by mixed-use developments put no strain on the external street network and should be deducted from ITE trip rates for standalone
suburban developments.
Trip distances are also variable across regions. For external auto
trips, average distances range from 4.6 miles for MXDs in Boston
to 13.9 miles for MXDs in Houston (Table 4). Again, this reflects
the size and remoteness of Houston’s MXDs.
Table 5 provides comparable data for trips internal to MXDs. In
four of the six regions, approximately half of the internal trips are
walk trips, and auto trips are short. For these regions, it may be
reasonable to ignore the contribution of internal trips to regional
VMT and emissions, particularly since internal trips are only
approximately 16% of all trips produced by or attracted to these
MXDs. For the remaining two regions, with their large masterplanned communities, the share of walk trips is low, and auto trips
are relatively long. For these regions, the contribution of internal
trips to regional VMT and emissions is significant. Although these
trips may not contribute to area-wide congestion, they should be
considered in VMT and emissions calculations.
252 / JOURNAL OF URBAN PLANNING AND DEVELOPMENT © ASCE / SEPTEMBER 2011
Downloaded 14 Sep 2011 to 192.245.136.3. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
Table 2. Variable Definition and Description
Outcome variables
INTERNAL
WALK
TRANSIT
TDIST
Explanatory variables
Definition
Dummy variable indicating that a trip remains internal to the MXD (1 ¼ internal, 0 ¼ external).
Dummy variable indicating that the travel mode on an external trip is walking (1 ¼ walk, 0 ¼ other).
Dummy variable indicating that the travel mode on an external trip is public bus or rail (1 ¼ transit, 0 ¼ other).
Network trip distance between origin and destination locations for an external private vehicle trip, in miles.
Level 1 traveler/household level
CHILD
HHSIZE
VEHCAP
BUSSTOP
Variable
Number
Number
Dummy
indicating that the traveler is under 16 years of age (1 ¼ child, 0 ¼ adult).
of members of the household.
of motorized vehicles per person in the household.
variable indicating that the household lives within 1=4 mile of a bus stop (1 ¼ yes, 0 ¼ no)
Level 2 MXD explanatory variables
AREA
POP
EMP
ACTIVITY
ACTDEN
DEVLAND
JOBPOP
LANDMIX
STRDEN
INTDEN
EMPMILE
EMP30T
EMP10A, EMP20A,
EMP30A
STOPDEN
RAILSTOP
Gross land area of the MXD in square miles.
Resident population within the MXD; prorated sum of the population for the census block groups that intersect the MXD. Prorating
was done by calculating density of population per residential acre (tax lots designated single-family or multifamily) for the entire
census block group, then multiplying the density by the amount of residential acreage within the block group contributing to the
MXD, and finally, summing over all block groups intersecting the MXD area. For Houston, data at the traffic analysis zone (TAZ)
level were prorated.
Employment within the MXD; weighted sum of the employment within the MXD for all Standard Industrial Classification (SIC)
industries. For Portland, employment estimates were based on the average number of employees in each size category, summed
across employer size categories. For other regions, data at the TAZ level were prorated.
Resident population plus employment within the MXD.
Activity density per square mile within the MXD. Sum of population and employment within the MXD, divided by gross land area.
Proportion of developed land within the MXD.
Index that measures balance between employment and resident population within MXD. Index ranges from 0, where only jobs or
residents are present in an MXD, not both, to 1 where the ratio of jobs to residents is optimal from the standpoint of trip generation.
Values are intermediate when MXDs have both jobs and residents, but one predominates.a
Another diversity index that captures the variety of land uses within the MXD. This is an entropy calculation based on net acreage in
land-use categories likely to exchange trips. For Portland, the land uses were: residential, commercial, industrial, and public or
semipublic.b For other regions, the categories were slightly different.c The entropy index varies in value from 0, where all developed
land is in one of these categories, to 1, where developed land is evenly divided among these categories.
Centerline miles of all streets per square mile of gross land area within the MXD.
Number of intersections per square mile of gross land area within the MXD.
Total employment outside the MXD within one mile of the boundary. Weighted average for all TAZs intersecting the MXD.
Weighting was done by proportion of each TAZ within the MXD boundary relative to an entire TAZ area (i.e., “clipping” the block
group with the MXD polygon).
Share of total regional employment accessible within 30-min travel time of the MXD using transit.
Share of total regional employment accessible within 10, 20, and 30-min travel time of the MXD using an automobile at midday.
Number of transit stops within the MXD per square mile of land area. Uses 25 ft buffer to catch bus stops on periphery.
Rail station located within the MXD (1 ¼ yes, 0 ¼ no). Commuter, metro, and light rail systems are all considered.
Level 3 regional explanatory variables
REGPOP
REGEMP
REGACT
SPRAWL
Population within the region.
Employment within the region.
Activity within the region (population þ employment).
Measure of regional sprawl developed by Ewing et al. (2002, 2003). Index derived by extracting the common variance from multiple
measures through principal components analysis.
a
JOBPOP ¼ 1 ½ABSðemployment 0:2 populationÞ=ðemployment þ 0:2 populationÞ; ABS is the absolute value of the expression in parentheses. The
value 0.2, representing a balance of employment and population, was found through trial and error to maximize the explanatory power of the variable.
b
The entropy calculation is LANDMIX ¼ ½single-family share LNðsingle family shareÞ þ multifamily share LNðmultifamily shareÞþ commercial share
LNðcommercial shareÞ þ industrial share LNðindustrial shareÞ þ public share LNðpublic shareÞ=LNð5Þ, where LN is the natural logarithm.
c
For Houston, the land uses were: residential, commercial, industrial, and institutional; a mixed residential and commercial class of land uses was included
with commercial. For Boston, the land uses were: residential, commercial, industrial, and recreational. For Seattle, detailed land uses were aggregated into
four categories: residential, commercial, industrial, and institutional. For Atlanta, detailed land uses were aggregated into four categories: residential,
commercial, industrial, and institutional. For Sacramento, detailed land uses were aggregated into four categories: residential, commercial, industrial,
and institutional; a mixed class of land uses was included with commercial.
JOURNAL OF URBAN PLANNING AND DEVELOPMENT © ASCE / SEPTEMBER 2011 / 253
Downloaded 14 Sep 2011 to 192.245.136.3. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
Table 3. Sample Sizes and Descriptive Statistics for Levels 1 and 2
Variables
INTERNAL
WALK
TRANSIT
TDIST
CHILD
HHSIZE
VEHCAP
BUSSTOP
AREA
POP
EMP
ACTIVITY
ACTDEN
DEVLAND
JOBPOP
LANDMIX
STRDEN
INTDEN
EMPMILE
EMP30T
EMP10A
EMP20A
EMP30A
STOPDEN
RAILSTOP
N
Mean
SD
35,877
29,499
29,499
23,921
35,877
35,877
35,877
35,877
239
239
239
239
239
239
239
239
239
239
239
239
239
239
239
239
239
0.18
0.07
0.07
6.48
0.10
2.66
0.80
0.43
0.33
2,271
2,696
4,967
19,780
0.83
0.46
0.52
25.4
257
30,510
0.058
0.048
0.185
0.336
70.8
0.08
0.38
0.26
0.25
7.79
0.30
1.32
0.47
0.50
0.32
3,261
5,572
6,945
30,669
0.22
0.31
0.20
10.5
203
50,914
0.095
0.073
0.230
0.391
83.6
0.28
Table 4. Average Internal Capture Rates, Walk and Transit Mode Shares
for External Trips, and Auto Trip Distances for External Trips to/from
MXDs
Region
Atlanta
Boston
Houston
Portland
Sacramento
Seattle
Overall
Mode share percentages
for external trips
Internal
capture
(percentage
of all trips)
Walk
share
Transit
share
16.7%
16.9%
31.1%
15.9%
16.4%
18.0%
17.8%
5.0%
20.6%
3.1%
7.3%
2.9%
5.8%
7.1%
3.1%
7.8%
6.1%
4.6%
0.4%
9.9%
6.8%
Auto
distance
Sum of walk for external
and transit trips (miles)
8.1%
28.4%
9.3%
11.9%
3.3%
15.7%
13.9%
6.3
4.6
13.9
4.8
6.8
6.9
6.5
Table 5. Walk and Transit Mode Shares and Auto Trip Distances for Trips
Internal to MXDs
Atlanta
Boston
Houston
Portland
Sacramento
Seattle
Overall
Walk share
of internal
trips
Transit share
of internal
trips
Sum of walk
and transit
internal trips
Auto distance
of internal
trips (miles)
53.7%
54.3%
15.0%
43.4%
7.4%
57.1%
47.7%
0.8%
1.0%
4.5%
0.8%
0%
1.1%
1.2%
54.5%
55.3%
19.5%
44.2%
7.4%
58.2%
48.9%
0.45
0.51
3.29
0.57
0.64
0.36
0.81
Models
As indicated, four outcomes are modeled in this study: choice of
internal destination, choice of walking on external trips, choice
of transit on external trips, and distance of external trips by private
vehicle. Models apply to both trips produced by and trips attracted
to MXDs. Models are estimated separately by trip purpose: homebased work, home-based other, and non-home-based. The writers
presume that different factors might be at play, or that the same
factors might be more or less important when people travel for
different purposes. Modeling by trip purpose also gives us some
ability to distinguish peak hour travel (disproportionately homebased work) from off-peak travel (disproportionately home-based
other and non-home-based).
The writers took an exploratory approach in modeling factors
that could explain outcome variables, seeking to include at least
one variable from each of the six Ds. To keep the results parsimonious and avoid possible multicollinearity problems, the threshold
for inclusion of variables in models was a significance level of 0.10.
A majority of variables included in our models have much higher
significance levels than this threshold value.
For internal capture, our dependent variable is the natural log of
the odds of an individual making a trip with both ends within an
MXD. For external walk and transit trips, the dependent variable is
the natural log of the odds of an individual making a trip by these
modes. For external private vehicle trips, the dependent variable is
the distance from origin to destination in miles.
For these outcomes, models have been estimated with both
linear and logarithmic (natural log) values of the independent
variables. The logarithmic models, which express the odds as a
power function of the independent variables, outperform the linear
models in terms of their pseudo-R2 s, sensitivity to changes in values of independent variables, and validation results (described in
the following). Thus, only the logarithmic models are presented
in this article. Coefficient values are arc elasticities of odds with
respect to the independent variables.
For estimating the trip distance by automobile, models took
three forms: linear, semilogarithmic (linear-log), and log-log forms.
The semilogarithmic models, which express trip distance as a linear
sum of logged variables, outperform the other models in terms of
their pseudo-R2 s and sensitivity to changes in values of independent variables. Only the semilogarithmic models are presented
in this article.
This study’s data and model structure are hierarchical. Thus,
hierarchical modeling is the best methodology to account for
dependence among observations, in this case the dependence of trips
to and from a given MXD and dependence of MXDs within a given
region. All the trips to/from a given MXD share the characteristics
of the MXD, that is, are dependent on these characteristics. This
dependence violates the independence assumption of ordinary
least squares (OLS) regression. Standard errors of regression coefficients based on OLS will consequently be underestimated.
Moreover, OLS coefficient estimates will be inefficient. Hierarchical
(multilevel) modeling overcomes these limitations, accounting
for the dependence among observations and producing more
accurate coefficient and standard error estimates (Raudenbush
and Bryk 2002).
The writers initially conceived the data structure as a five-level
hierarchy, with trips nested within individuals, individuals nested
within households, households nested within MXDs, and MXDs
nested within metropolitan regions. Upon review of the data set,
we found that the data are not so neatly hierarchical. Many of
the individuals in the sample make trips to or from more than
one MXD.
254 / JOURNAL OF URBAN PLANNING AND DEVELOPMENT © ASCE / SEPTEMBER 2011
Downloaded 14 Sep 2011 to 192.245.136.3. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
Results
Internal Capture
Fig. 3. Data and model structure
This has implications for modeling methodology. Rather than a
five-level hierarchy, the choices facing travelers have been modeled
in a three-level framework. Individual trip ends are uniquely identified with MXDs. Therefore, trips (their characteristics and the
associated characteristics of travelers and their households) form
Level 1 in the hierarchy, MXDs form Level 2, and regions form
Level 3 (Fig. 3). Within a hierarchical model, each level in the data
structure is formally represented by its own submodel. The submodels are statistically linked.
Models were estimated with hierarchical linear and nonlinear
modeling (HLM) 6 software. Hierarchical linear models were estimated for the continuous outcome (trip distance), and hierarchical
nonlinear models were estimated for the dichotomous outcomes
(internal versus external, walk versus other, and transit versus
other). Hierarchical linear modeling is analogous to linear
regression analysis, although models are estimated by using maximum likelihood estimation rather than OLS. Hierarchical nonlinear
modeling is analogous to logistic regression. Like logistic regression, hierarchical nonlinear modeling uses maximum likelihood
estimation.
In the initial model estimations, only the intercepts were allowed
to randomly vary across higher level units. All of the regression
coefficients at higher levels were treated as fixed. These are referred
to as random intercept models (Raudenbush and Bryk 2002).
As the sample of MXDs expanded, we also tested for cross-level
variable interactions with random coefficient models. It is certainly
possible that the relationship between, for example, walking and
vehicle availability varies with size of the MXD, or the relationship
between internal capture and MXD density varies from region
to region. As the cross-level interaction terms seldom proved
significant, only the random intercept models are presented in
the following section.
For internal capture of trips, coefficients and their significance
levels (p-values) are shown in Table 6. The coefficients are elasticities of the odds of internal capture with respect to the various
independent variables, that is, measures of effect size. In the case of
home-based work trips, the odds of an internal trip decline with
household size and vehicle ownership per capita, and increase with
an MXD’s job-population balance. Internal capture is thus related
to two D variables, diversity and demographics. Larger households
have more complex activity patterns that are more likely to take
them beyond the bounds of an MXD. Households with higher
vehicle ownership have fewer constraints on the use of household
vehicles for long trips. A high job-population balance value translates into more opportunities to live and work on-site. The pseudoR2 of this model is quite low, at 0.01, indicating a considerable
amount of unexplained influence on the odds of a home-based
work trip being internally captured.
For home-based other trips, the odds of internal capture decline
with household size and vehicle ownership per capita and increase
with an MXD’s land area, job-population balance, and intersection
density. Internal capture for trips from home to nonwork destinations is thus related to development scale, diversity, design, and
demographics. Relationships to household size and vehicle ownership are as explained previously. As for the other significant variables, job-population balance spawns on-site travel because jobs
include those in the retail sector, suggesting the presence of shops
and restaurants encourages some residents to substitute walk trips
for out-of-neighborhood car trips. Also, a large land area increases
the likelihood that nonwork destinations will be on-site while high
intersection density increases routing options, makes routes more
direct, creates frequent street crossing opportunities, and makes
trips seem more eventful. Among the built environment variables
analyzed, the most statistically significant predictor of internal
capture for home-based other trips is job-population balance followed by an MXD’s land area and then by its intersection density.
The pseudo-R2 of this model is a more respectable 0.20, but still
indicates that there is a considerable amount of unexplained influence on the odds of a trip being internally captured
For non-home-based trips, the odds of internal capture decline
with household size and vehicle ownership, and increase with land
area, employment, and intersection density of the MXD. Internal
capture is thus related to design, development scale, and demographics. Relationships to land area and intersection density were
explained previously. The relationship to employment is likely
attributable to the greater likelihood of matching employees’
Table 6. Log Odds of Internal Capture (Log-Log Form)
Home-based work
Coefficient
Constant
EMP
AREA
JOBPOP
INTDEN
HHSIZE
VEHCAP
Pseudo-R2
1:75
—
—
0.389
—
1:33
0:990
t-ratio
—
—
2.62
—
6:03
4:15
0.01
Home-based other
p-value
Coefficient
—
—
0.010
—
< 0:001
< 0:001
2:43
—
0.486
0.399
0.385
0:867
0:590
t-ratio
—
3.61
4.55
1.92
13:0
8:19
0.20
Non-home-based
p-value
Coefficient
—
0.001
< 0:001
0.055
< 0:001
< 0:001
5:32
0.208
0.468
—
0.638
0:237
0:163
t-ratio
3.28
4.58
—
4.95
4:54
3:00
0.30
p-value
0.002
< 0:001
—
< 0:001
< 0:001
0.003
JOURNAL OF URBAN PLANNING AND DEVELOPMENT © ASCE / SEPTEMBER 2011 / 255
Downloaded 14 Sep 2011 to 192.245.136.3. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
commutes. This is buttressed by the fact that the coefficient of
EMPMILE is positive, indicating that when off-site job opportunities are nearby, MXD residents will walk to work. The pseudo-R2
of this model is 0.19.
For external home-based other trips, the odds of walking decline
with household size and vehicle ownership per capita, decline with
the land area of the MXD, and increase with the activity density of
the MXD, the job-population balance within the MXD, and number
of jobs outside the MXD within a mile of the boundaries. Walking
on external trips is thus related to measures of development scale,
density, diversity, destination accessibility, and demographics. The
larger the area of the MXD, the longer the external trips and the less
likely they will be made by walking. The higher the activity density,
the better the pedestrian environment and the more accessible
attractions will be to those traveling into the community. Relationships to job-population balance and employment within a mile have
already been discussed. The pseudo-R2 of this model is a high 0.51.
For external non-home-based trips, the odds of walking decline
with household size and vehicle ownership per capita, and increase
with the activity density of the MXD, the intersection density of the
MXD, and the number of jobs outside the MXD within a mile of
the boundaries. Walking on external trips is thus related to measures
of density, design, destination accessibility, and demographics. High
intersection density within the MXD makes walking to/from activities outside the MXD that much more direct. The other independent
variables have already been discussed. The pseudo-R2 of this model
is a very high 0.64. Overall, the external walk models have the greatest explanatory power of all models estimated.
Table 8 shows the coefficients and significance levels (p-values)
of estimated models for predicting transit mode choice on external
trips. For external home-based work trips, the odds of transit use
decline with household size and vehicle ownership per capita, and
desired trips to on-site destinations when there are more attractions
nearby. The most statistically significant relationship is to intersection density, followed by land area, and then by employment. The
pseudo-R2 of this model is highest of the three trip purposes, 0.30.
Although there is significant variance of internal capture from
region to region, it is not explained by the variables in our data set.
None of the Level 3 variables proved significant. This is not too
surprising, given the small sample (six) of Level 3 units. Nonetheless, regional variance is captured in the random effects term of
the Level 3 equation, just not explained by any of the regional
variables.
Mode Choice for External Trips
Table 7 shows the coefficients and significance levels (p-values) of
estimated models for predicting walk mode choice on external trips.
The coefficients are elasticities of the odds of walking with respect
to the various independent variables, that is, measures of effect size.
For external home-based work trips, the odds of walking decline
with household size and vehicle ownership per capita, and increase
with job-population balance within the MXD and number of
jobs outside the MXD within a mile of the boundaries. Walking on
external trips is thus related to three types of D variables: diversity,
destination accessibility, and demographics. Large households
achieve economies through car pooling and trip chaining, and thus
are less likely to walk. Households with more cars have a lower
generalized cost of auto use, making them less likely to walk. Reasons for the positive association between internal job-population
balance and walking for external work trips are less obvious.
One possibility is that on-site balance creates opportunities for trip
chaining. Another possibility is that on-site balance is associated
with off-site, nearby balance as well, thus further inducing walk
Table 7. Log Odds of Walking on External Trips (Log-Log Form)
Home-based work
Coefficient
Constant
AREA
ACTDEN
JOBPOP
INTDEN
EMPMILE
HHSIZE
VEHCAP
Pseudo-R2
t-ratio
Home-based other
p-value
5:55
Non-home-based
Coefficient
t-ratio
p-value
4:27
2.74
3.83
< 0:001
0.007
< 0:001
5.05
5:05
7:62
0.51
0.226
2.46
0.015
10:96
0:415
0.370
0.219
0.385
1:57
1:84
3.12
6:29
7:00
0.19
0.002
< 0:001
< 0:001
0.450
0:486
0:768
Coefficient
t-ratio
p-value
0.377
3.12
0.003
< 0:001
< 0:001
< 0:001
0.803
0.440
0:281
0:242
5.05
5.09
2:59
2:13
0.64
< 0:001
< 0:001
0.010
0.033
t-ratio
p-value
Coefficient
2.89
0.005
8:48
8:91
4.04
NA
< 0:001
< 0:001
< 0:001
15:09
Table 8. Log Odds of Using Transit on External Trips (Log-Log Form)
Home-based work
Coefficient
Constant
ACTDEN
INTDEN
EMP30T
HHSIZE
VEHCAP
BUSSTOP
Pseudo-R2
t-ratio
Home-based other
p-value
8:05
1.12
0.209
1:14
1:68
0.357
Coefficient
6:08
0.324
4.44
2.98
6:31
8:56
2.08
0.47
< 0:001
0.004
< 0:001
< 0:001
0.037
0:958
1:09
0.467
Non-home-based
t-ratio
p-value
0.134
3.29
0.002
0:340
3:74
< 0:001
2:69
NA
256 / JOURNAL OF URBAN PLANNING AND DEVELOPMENT © ASCE / SEPTEMBER 2011
Downloaded 14 Sep 2011 to 192.245.136.3. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
increase with the intersection density of the MXD and the number
of jobs within a 30-min trip by transit. The odds of transit use are
significantly higher for households living within 1=4 mile of a bus
stop than those farther away. Transit use on external trips is thus
related to measures of design, destination accessibility, distance to
transit, and demographics. A higher intersection density translates
into a more direct walk trips to and from transit stops, and also
possibly more efficient routing of transit vehicles. More jobs within
30 min by transit increase the likelihood a particular job being
within easy commuting distance for residents. Residence within
the standard quarter mile walking distance of a bus stop shortens
access trips. The pseudo-R2 of this model is 0.48.
For external home-based other trips, the odds of transit use
decline with household size and vehicle ownership per capita and
increase with the activity density within the MXD. The odds are
significantly higher for households living within 1=4 mile of a
bus stop than those further away. The higher the activity density,
the better the pedestrian environment and the more accessible
attractions will be to those traveling into the community. The other
independent variables have already been discussed. This is a weak
model. The pseudo-R2 of this model is a negative number because
the combined variance at Levels 1 through 3 is greater for the
estimated model than the null model with only an intercept and
no explanatory variables.
For external non-home-based trips, the odds of transit use
decline with household size and vehicle ownership per capita,
and increase with the number of jobs within a 30 min trip by transit.
These independent variables have already been discussed. The
pseudo-R2 of this model also is a negative number.
Regarding these negative pseudo-R2 s, a pseudo-R2 is not entirely analogous to R2 in linear regression, which can only assume
positive values. One standard text on multilevel modeling notes that
the variance can increase when variables are added to the null
model. It goes on to say: “This is counterintuitive, because we have
learned to expect that adding a variable will decrease the error variance, or at least keep it at its current level… In general, we suggest
not setting too much store by the calculation of [pseudo-R2s]”
(Kreft and de Leeuw 1998). For more discussion of negative
pseudo-R2 s, also see Snijders and Bosker (1999).
Activity density has the expected positive sign in all three regressions. It reaches statistical significance in only one regression.
This is consistent with a finding from a recent meta-analysis of the
built environment-travel literature that density is the least important
of the D variables (Ewing and Cervero 2010). Having a rail stop
within a development also has a positive sign in all three regressions but never reaches statistical significance.
Although there is significant variance of walking and transit
use from region to region, it is not explained by the variables in
our data set. Again, none of the Level 3 variables proved significant. Regional variance is, however, captured in the random effects
term of the Level 3 equation.
Trip Distance for External Automobile Trips
Table 9 shows the coefficients and significance levels (p-values) of
estimated models for predicting auto trip distances on external trips.
For external home-based work trips by private vehicle, trip distance
increases with household size, vehicle ownership per capita, and
land area of the MXD, and declines with a project’s job-population
balance and the share of regional jobs reachable within 30 min by
automobile. External trip length is thus related to four types of D
variables, development scale, diversity, destination accessibility,
and demographics. Larger MXDs produce and attract longer external trips simply because the shortest trips are internalized. MXDs
with good job-population balance apparently reduce the need for
very long external trips; e.g., on-site residents patronizing on-site
retail outlets. They may also facilitate trip chaining. MXDs with
good auto accessibility to regional jobs generate shorter trips because more trip attractions are nearby. On the other hand, larger
households have more complex activity patterns, which lengthen
trips. More vehicles per household frees up family cars for trips
to more distant destinations. These relationships match expectations. The pseudo-R2 is 0.11.
For external home-based other trips, trip distance increases with
household size and vehicle ownership per capita, and declines with
the job-population balance within the MXD and the share of
regional jobs reachable within 20 min by automobile. External
trip length is thus related to measures of diversity, destination
accessibility, and demographics. Relationships to job-population
balance, accessibility to regional employment, household size,
and vehicle ownership follow the same explanations provided
above. The destination accessibility measure with the greatest
explanatory power is the number of jobs reachable within
20 min by automobile, not 30 min as with home-based work trips.
This makes sense, because home-based other trips are shorter than
home-based work trips. The pseudo-R2 of this model is 0.03.
For external non-home-based trips, trip distance increases with
household size and vehicle ownership per capita, and declines with
the job-population balance within the MXD, intersection density
within the MXD, and the share of regional jobs reachable within
20 min by automobile. External trip length is thus related to
measures of diversity, design, destination accessibility, and demographics. As for the one new variable, higher intersection density
within an MXD (and perhaps its surroundings as well) makes for
Table 9. Trip Distance for External Automobile Trips (Semilog Form)
Home-based work
Constant
AREA
JOBPOP
INTDEN
EMP20A
EMP30A
HHSIZE
VEHCAP
Pseudo-R2
Home-based other
Coefficient
t-ratio
p-value
6.54
1.07
0:298
2.92
1:88
0.004
0.061
6:05
8.08
7.26
0.11
< 0:001
< 0:001
< 0:001
1:19
2.76
2.76
Coefficient
t-ratio
Non-home-based
p-value
4.33
Coefficient
t-ratio
p-value
2:05
2:06
5:69
0.041
0.041
< 0:001
2.58
5.12
0.05
0.010
< 0:001
8.99
0:356
2:38
0.018
0:697
4:79
< 0:001
0:282
0:832
0:823
0.772
1.48
5.06
9.22
0.03
< 0:001
< 0:001
0.520
1.06
JOURNAL OF URBAN PLANNING AND DEVELOPMENT © ASCE / SEPTEMBER 2011 / 257
Downloaded 14 Sep 2011 to 192.245.136.3. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
more direct connections to external trip attractions. The pseudo-R2
of this model is 0.05.
The VMT calculations made possible with the models presented
in Tables 6–9 represent only the mileage generated by travel external to the development site. For large MXDs where internal vehicle
travel is likely, users of the MXD method are advised to perform an
independent estimate of internal vehicle trip generation and VMT.
In these cases, internal and external VMT should be combined for
complete estimates of impacts such as fuel consumption, greenhouse gases, and other emissions.
Model Validation
For this method to gain credibility, it is important that the results be
validated by comparing estimates to in-field traffic counts. The preceding models were applied to 22 MXDs for which traffic counts
of external vehicle trips were available. Six of those 22 sites are
located in South Florida. Their traffic counts are presented in
Appendix C of the Trip Generation Handbook (2004). Four additional sites are located in Central Florida, Atlanta, and Texas, of
which three are nationally known examples of smart growth or
transit-oriented development: Celebration Florida, Atlantic Station,
and Mockingbird Station. Six sites are located in San Diego County
and were designated by local planners and traffic engineers in 2009
as representing a wide range of examples of smart growth trip
generators. The six remaining sites are conventional development
projects located elsewhere in California. The sites represent a
wide range of densities, land-use mixes, and development scales.
Populations of the validation MXDs range from zero (Crocker
Center in Boca Raton, FL and Hazard Center in San Diego, containing a mix of commercial and office uses only) to nearly 17,000
(the entire town of Moraga, CA). Employment levels range from
near-zero (The Villages in Irvine, CA, which is predominantly
residential, with only a small amount of restaurant and service retail) to more than 5,500 (Park Place, also in Irvine, CA). Some sites
are well served by transit, including three built around rail stations,
whereas others are suburban and poorly served by transit. With
such a diverse validation sample, one can begin to build confidence
that these MXD models have external validity.
Data were collected for all model variables at each of the 22
sites. The variables EMPMILE and EMP30T were estimated from
regional travel models for the MXD traffic analysis zones and visually verified from aerials, and in some cases, from websites of the
MXDs. For those sites for which household data were not available,
the household size and vehicle ownership variables for trips
produced and attracted to the MXD were taken from 2000 census
data for the census tracts most closely matching the locations of
the MXDs.
The probabilities estimated with these models and the resulting
predicted external vehicle traffic counts are shown in Table 10. The
results demonstrate that the models are capable of predicting a wide
range of internal capture rates and mode shares for external trips,
taking into account development scale, site design, and regional
context. The models predict total vehicle counts within 20% of
the actual number of trips observed for 15 of the 22 validation sites,
within 30% for four sites, and within 40% for another one. Only
two sites were off by more than 40%. When compared with the best
available published methods for estimating trip generation, the
models improved the prediction of vehicle counts at 16 of the
22 validation sites.
Table 11 compares model performance to current methods,
specifically:
1. ITE Trip Generation or SANDAG Traffic Generators (2004)
without any adjustments (Gross trips); and
2. Current internalization methods from the ITE Trip Generation
Handbook or from SANDAG’s current method of deducting
5% for mixed-use and 10% for proximity to transit (Net trips)
Table 10. Predicted Probabilities from Application of the Model to Validation Sites
Site name and location
Atlantic Station, Atlanta, GA
Boca Del Mar, Boca Raton, FL
Town of Celebration, Celebration, FL
Country Isles, Weston, FL
Crocker Center, Boca Raton, FL
Galleria, Ft. Lauderdale, FL
Gateway Oaks, Sacramento, CA
Jamboree Center, Irvine, CA
Legacy Town Center, Plano, TX
Mizner Park, Boca Raton, FL
Mockingbird Station, Dallas, TX
Town of Moraga, Moraga, CA
Park Place, Irvine, CA
South Davis, Davis, CA
The Villages, Irvine, CA
Village Commons, West Palm Beach, FL
Rio Vista, San Diego, CA
Village Plaza, La Mesa, CA
Uptown Center, San Diego, CA
Morena Linda Vista, San Diego, CA
Hazard Center, San Diego, CA
Otay Ranch, Chula Vista, CA
Internal
capture rate
External walk
mode share
External transit
mode share
Predicted external
vehicle counts
Observed external
vehicle counts
11%
10%
26%
9%
3%
9%
11%
9%
14%
8%
6%
28%
7%
25%
2%
6%
4%
7%
6%
6%
5%
5%
7%
3%
2%
0%
4%
3%
4%
5%
12%
9%
19%
1%
7%
2%
7%
4%
15%
17%
17%
16%
10%
3%
4%
2%
0%
0%
3%
3%
3%
4%
4%
5%
6%
1%
4%
2%
4%
0%
7%
8%
6%
7%
8%
4%
31,377
20,890
35,775
14,891
17,077
29,505
16,320
36,039
24,903
11,559
11,153
55,816
17,417
66,752
7,680
22,793
5,024
3,920
14,734
4,132
11,685
9,279
28,787
22,846
40,912
22,419
9,791
22,971
23,280
36,569
20,082
12,086
20,677
49,689
19,064
74,648
7,128
18,075
5,307
4,280
16,886
4,712
11,644
7,935
258 / JOURNAL OF URBAN PLANNING AND DEVELOPMENT © ASCE / SEPTEMBER 2011
Downloaded 14 Sep 2011 to 192.245.136.3. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
Table 11. Comparison of Percent Differences between Predicted and
Observed External Vehicle Counts by Gross, Net, and MXD Modeling
Methods
Site name and location
Gross
tripsa
Atlantic Station, Atlanta, GA
Boca Del Mar, Boca Raton, FL
Town of Celebration, Celebration, FL
Country Isles, Weston, FL
Crocker Center, Boca Raton, FL
Galleria, Ft. Lauderdale, FL
Gateway Oaks, Sacramento, CA
Jamboree Center, Irvine, CA
Legacy Town Center, Plano, TX
Mizner Park, Boca Raton, FL
Mockingbird Station, Dallas, TX
Town of Moraga, Moraga, CA
Park Place, Irvine, CA
South Davis, Davis, CA
The Villages, Irvine, CA
Village Commons, West Palm Beach, FL
Rio Vista, San Diego, CA
Village Plaza, La Mesa, CA
Uptown Center, San Diego, CA
Morena Linda Vista, San Diego, CA
Hazard Center, San Diego, CA
Otay Ranch, Chula Vista, CA
RMSE
R2
37%
7%
21%
27%
94%
50%
15%
20%
74%
20%
24%
59%
10%
24%
22%
40%
26%
33%
20%
35%
29%
32%
44%
0.65
MXD
Net tripsb method
25%
4%
17%
38%
82%
35%
19%
8%
50%
20%
26%
49%
4%
3%
17%
31%
8%
13%
8%
16%
11%
19%
32%
0.81
9%
9%
13%
34%
74%
28%
30%
1%
24%
4%
46%
12%
9%
11%
8%
26%
5%
8%
13%
12%
0%
17%
22%
0.92
a
Gross trips estimates are computed from the trip generation rates contained
in the ITE Trip Generation report (sites 1–16) or the SANDAG Traffic
Generators report (sites 17–22) for each of the individual land uses within
the mixed-use site, without discounting for internalization, walking or
transit use.
b
Net trips estimates apply internalization reductions for multiuse sites
as prescribed in the ITE Trip Generation Handbook (sites 1–16) or the
SANDAG Traffic Generators report (sites 17–22) and represent the best
estimates that one could obtain relying on currently published material
alone.
Percentages reported in Table 11 indicate errors in trip estimates
using gross, net, and MXD modeling methods, respectively, for
each testing site. The percent root mean squared error (%RMSE),
used in the transportation field to evaluate model accuracy, penalizes proportionally more for large errors and normalizes the error
across different values of the quantity one is trying predict. A %
RMSE of less than 40% is generally considered good. Table 11
shows that the proposed models improve the %RMSE over the
gross and net methods. The MXD models improve the %RMSE
from 32%, produced by the best of the previously available trip
generation estimation methods, to a figure of 22%.
R2 in this table measures the squared difference between the observed and predicted external vehicle counts as a percentage of the
squared variation of the observed external vehicle counts about the
mean over the 22 sites. Table 11 shows that the proposed models
also improve the R2 significantly compared to the gross and net
methods. The R2 for the MXD model is 0.92, markedly better than
the 0.81 value for the net method, the best estimates that one could
obtain relying on previously published material alone.
Finally, Figs. 4 and 5 show the strong association between
predicted and observed external vehicle counts using the models
Fig. 4. Scatterplot of predicted versus observed external vehicle counts
developed herein, and a comparison of daily observed external
vehicle counts across the three methods.
Applications
The previously derived models can be used to predict trip productions plus attractions for three separate trip purposes. Having the
models for three trip purposes allows the practitioner to predict
external private vehicle trips on either a daily basis or a morning
or afternoon peak hour basis. The likelihood that a trip during any
of these times of day is home-based-work or home-based-other or
non-home-based is determined based on purpose-specific trip generation rates in NCHRP Report 365, Travel Estimation Techniques
for Urban Planning (1998). The log odds estimates of internal
capture or walk or transit, as obtained from Tables 6–8, are first
exponentiated, then converted into probabilities using the formula:
probability ¼ odds=ð1 þ oddsÞ. They are then applied to each
estimate of total trips generated for the three trip purposes. The
remaining trips are combined across all three trip purposes to get
total net external private vehicle trips.
The models are applied in sequential fashion for each trip
purpose. The probability of trips for a given trip purpose traveling
external to the site is computed first, using the equations in Table 6.
These resulting probabilities are used to discount the total site trip
generation as estimated by the trip rates contained in the ITE Trip
Generation report. The resulting external trips are then further
reduced to account for those external trips that would probabilistically travel by walking or transit, using the equations provided in
Tables 7 and 8. The three trip purposes are then combined. This
leaves the estimate of external vehicle trips, or in ITE terms, site
traffic generation. For those who want to compute the VMT generated by the site, one would apply the Table 9 equations to compute the average external trip lengths, and for large sites, would add
an estimate of VMT generated by trips remaining entirely within
the site.
Most of the information required to apply these equations is
readily available from project site plans and programmatic data developed as part of a project planning process or submitted as part of
a development application. Certain data items may require the traffic engineer to obtain information via GIS mapping of the site area
or a request to the jurisdiction or metropolitan planning organization for information easily extracted from the regional travel model.
These data, when incorporated into equations that estimate MXD
internal capture and walking and transit use, produce trip generation reductions to be applied to estimates based on the ITE Trip
Generation report.
JOURNAL OF URBAN PLANNING AND DEVELOPMENT © ASCE / SEPTEMBER 2011 / 259
Downloaded 14 Sep 2011 to 192.245.136.3. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
Fig. 5. (Color) Comparison of external vehicle trips across methods
As is the case with many of the guidelines presented in Trip
Generation, expert judgment is advised on case-by-case basis. This
might be necessary if, for example, in the judgment of a qualified
traffic engineer or planner, the development proposal under study is
unique in its relative composition of restaurant, theater, or other
commercial uses.
Conclusion
The bibles of traffic impact analysis, the Institute of Transportation
Engineers’ Trip Generation report (2008) and Trip Generation
Handbook (2004), are sorely lacking when it comes to MXDs.
Except for a handful of master-planned projects in Florida, actual
studies of internal capture rates are few and far between. Traffic
engineers are thus largely left to their own devices when quantifying the trip reductions that might result from mixing land uses.
Therefore, to err on the conservative side and avoid possible liability charges from underdesigning road capacity, often no adjustment
is made at all. This results in overestimates of the traffic impacts of
MXD proposals, leading to higher development fees than necessary
and raising opposition among those who fear potential adverse
impacts. Failure to account for internal capture and external walk
and transit trips ends up penalizing MXDs and can force MXD
developers to, in effect, cross-subsidize single-use projects through
disproportionate exactions. In addition, lack of accounting for
the trip-reducing benefits of MXDs can result in an oversupply
of parking.
This research sought to advance the state of knowledge on the
relationships that govern travel to, from, and within mixed-use development projects and to enumerate tangible and verifiable traffic
reductions relative to the rates in the ITE Trip Generation report.
Travel research published over the last few years convincingly
shows that changes by several percentage points in any or several
of the D variables used in this study reduces the number of vehicle
trips and vehicle miles traveled (Ewing and Cervero 2001, 2010).
This study extends and focuses that research on the particular
characteristics of MXDs. It represents the first national study of
the traffic generation by mixed-use developments, making use
of household travel survey data from six metropolitan regions.
The writers found that an average of three out of 10 trips generated
by MXDs put no strain on the external street network and generate
relatively few vehicle miles traveled. Statistical equations derived
from the data reveal that the primary factors affecting this reduction
in automobile travel are:
1. The total and the relative amounts of population and employment on the site;
2. The site size and activity density;
3. The size of households and their auto ownership;
4. The amount of employment within walking distance of the site;
5. The block size on the site; and
6. The access to employment within a 30 min transit ride of
the site.
For traffic impact, greenhouse gas, and energy analyses, the VMT
generated by a mixed-use site depends, in addition to the previously described factors, on the site’s placement within the region,
specifically, on the share of jobs located within a 20- to 30-min
drive of the site. Greater destination accessibility translates into
shorter auto trips external to the site. This effect is as significant
as the effects associated with internal capture of trips within
mixed-use developments, and conversion of some external trips
from auto to alternate modes.
This study’s findings regarding the factors that influence mixeduse trip generation have been validated through field surveys at
illustrative sites in California, Florida, Georgia, and Texas. The results will help guide planners and developers of mixed-use projects
on design features likely to minimize traffic generation, greenhouse
gas emissions, and energy impacts, and will produce new analysis
techniques for traffic engineers to more realistically quantify infrastructure impacts of mixed-use development proposals.
There are five caveats for practitioners. First, although MXDs
offer the option of walking, not all internally captured trips are walk
trips. This study focuses on MXD effects on external trip generation. Microscale built environmental features and their influence on
short-distance driving and nonmotorized trip-making in MXDs
warrant further investigation.
Second, when applying these models, internal capture rates
computed with the formulas are presumptive rates. They still need
to be adjusted to balance productions and attractions within the site,
as with the ITE Trip Generation Handbook method.
260 / JOURNAL OF URBAN PLANNING AND DEVELOPMENT © ASCE / SEPTEMBER 2011
Downloaded 14 Sep 2011 to 192.245.136.3. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org
Third, owing to limitations of the hierarchical modeling software the writers used (HLM), we specified two binomial mode
choice models (walk or not, transit or not) rather than one multinomial mode choice model (walk, transit, or auto). As a suggestion
for future research, it might be possible to estimate a multinomial
model with different software. As it is, the “not” alternative in each
case (“not walking,” “not transit”) is quite heterogeneous, including
all other alternatives. It would be more behaviorally sound (and
therefore may well increase the goodness-of-fit of the models)
to explicitly divide the “not” alternative into its constituent modes:
a traveler probably does not usually say, “should I take transit or
not?” but rather, “should I drive, walk, or take transit?”
Fourth, while acknowledging that walk trips may supplement
rather than substitute for private vehicle trips, we have in our validation exercise treated walking and transit use as one-for-one, tripfor-trip substitutes for private vehicle trips. Our data set prevents us
from estimating trip generation rates by mode because we have only
a sample of trips to, from, and within MXDs to work with, not the full
set of trip ends for nonresidential trip generators. This, in turn, forces
us to estimate mode choice equations, and keeps us from drawing
any inference about trip substitution. Because some of the walk trips
may supplement automobile trips, our walk mode choice models
represent an upper bound on actual rates of substitution.
Finally, the MXD methods presented here do not explicitly account for the effects of parking supply and price. Although several
of the variables included in the analysis, such as development density and proximity to regional job centers, may be partial proxies
for parking supply and price, site-specific parking data were not
available and were, therefore, not included in the MXD models.
If parking supply and price data were available they might significantly improve the ability to predict trip generation. It would also
inform the discussion of how MXDs can reduce trip generation by
pricing, supply constraint and unbundling the cost of parking from
the cost of real estate.
In closing, smart growth requires smart calculations. Unless
developers are rewarded for the trip-reducing impacts of MXDs,
the market incentive to build projects with relatively small environmental footprints is substantially reduced. Although the technical
aspects of this work might not be accessible to city planning commissioners and lay citizens, the basic premise that good development should be rewarded can be understood by all.
References
Boarnet, M. G., and Crane, R. (2001). “The influence of land use on travel
behavior: Specification and estimation strategies.” Transp. Res. Part A,
35(9), 823–845.
Cao, X., Mokhtarian, P. L., and Handy, S. L. (2009). “The relationship between the built environment and non-work travel: A
case study of northern California.” Transp. Res. Part A, 43(5),
548–559.
Cervero, R. (1988). “Land-use mixing and suburban mobility.” Transportation Quarterly, 42(3), 429–446.
Cervero, R. (2002). “Induced travel demand: Research design, empirical
evidence, and normative policies.” J. Plann. Lit., 17(1), 3–20.
Cervero, R., and Arrington, G. B. (2008). “Vehicle trip reduction impacts
of transit-oriented housing.” Journal of Public Transportation, 11(3),
1–18.
Cervero, R., and Kockelman, K. (1997). “Travel demand and the
3Ds: Density, diversity, and design.” Transp. Res. Part D, 2(3),
199–219.
Crane, R. (1996). “On form versus function: Will the new urbanism reduce
traffic, or increase it?” J. Plann. Educ. Res., 15, 117–126.
Ewing, R., and Cervero, R. (2001). “Travel and the built environment:
A synthesis.” Transp. Res. Rec., 1780, 87–113.
Ewing, R., and Cervero, R. (2010). “Travel and the built environment:
A meta-analysis.” J. Am. Plann. Assoc., 76(3), 265–294.
Ewing, R., Dumbaugh, E., and Brown, M. (2001). “Internalizing travel by
mixing land uses: Study of master-planned communities in south
Florida.” Transp. Res. Rec., 1780, 115–120.
Ewing, R., Pendall, R., and Chen, D. (2002). Measuring sprawl and its
impact, Smart Growth America/USEPA, Washington, DC.
Ewing, R., Pendall, R., and Chen, D. (2003). “Measuring sprawl and its
transportation impacts.” Transp. Res. Rec., 1832, 175–183.
Kreft, I., and de Leeuw, J. (1998). Introducing multilevel modeling, Sage
Publications, Thousand Oaks, CA, 115–119.
National Cooperative Highway Research Program (NCHRP). (1998).
“Travel estimation techniques for urban planning.” Rep. 365, Transportation Research Board and National Academy Press, Washington, DC,
82–90.
Raudenbush, S. W., and Bryk, A. S. (2002). Hierarchical linear models:
Applications and data analysis methods, 2nd Ed., Sage Publications,
Thousand Oaks, CA.
San Diego Association of Governments (SANDAG). (2004). San Diego
traffic generators, San Diego, CA.
Shoup, D. (2003). “Truth in transportation planning.” J. Transp. Stat., 6(1),
1–12.
Snijders, T., and Bosker, R. (1999). Multilevel analysis: An introduction to
basic and advanced multilevel modeling, Sage Publications, Thousand
Oaks, CA, 99–105.
Institute of Transportation Engineers (ITE). (2004). Trip generation
handbook, 2nd Ed., ITE, Washington, DC.
Institute of Transportation Engineers (ITE). (2008). Trip generation,
8th Ed., ITE, Washington, DC.
Zhang, M. (2004). “The role of land use in travel mode choice: Evidence
from Boston and Hong Kong.” J. Am. Plann. Assoc., 70(3),
344–361.
JOURNAL OF URBAN PLANNING AND DEVELOPMENT © ASCE / SEPTEMBER 2011 / 261
Downloaded 14 Sep 2011 to 192.245.136.3. Redistribution subject to ASCE license or copyright. Visithttp://www.ascelibrary.org